news

author｜song wanxin

editor｜zheng huaizhou

china has more than 7 million programmers. in the past few decades, the programming work of this large group has been almost entirely based on programming tools and bases provided overseas.

but that has changed in the past year.according to csdn data, 64.3% of ai programming tools are now domestically produced, and the use of domestic ai programming tools accounts for 57.2%.

thanks to the arms race among large domestic model manufacturers, alibaba cloud's ai programming tool "tongyi lingma" is now second only to the leading u.s. company github in terms of usage share, and more domestic tool products are also in the forefront. .

what you need to know is that at the beginning of last year, the market for programming tools was still blank in the country. in the process of domestic large-scale model melee, programming ai has also become one of the important competitive directions.

supply-side and demand-side growth are synchronous. jiang tao, chairman of csdn, said that the number of csdn users has increased by 7 million in the past year, and the proportion of developers paying attention to open source ai has reached 47%. the rapid growth of ai developers has become a significant trend.

"there are more than 1,000 basic open source models and domain open source models around the world, and the number of downloads of open source models in one year reached 350 million times. this is an era of ten times speed."

but for programmers, a large model that only "generates code" is not enough. software programming is a system that is complex enough to rival the automotive supply chain, but its degree of domestic industrialization is far less than that of the automotive industry chain.

after csdn and huawei cloud codearts released the open source community gitcode last year, on september 21 this year, the two parties collaborated to release the gitcode ai model community.

“gitcode ai model community benchmarks huggingface. now the developer core needs two bases, one for the code and the second for the data and models.”

jiang taozai36 kryptonin an interview, he said that generating code for large models is a capability. how to use this capability into its processes and coding tools is what csdn needs to do now.

according to the four stages of information technology development, the first-generation pc ecosystem, the second-generation internet ecosystem, and the third-generation operating system and chips, china and the world have maintained openness and interconnection in the first three generations.

but nowgeopoliticsunder the influence of the fourth-generation large-scale model and mobile phone intelligent application ecosystem, the domestic market needs to establish an independent and complete developer productivity tool system.

jiang tao believes that the fourth generation ecology is a major development opportunity for china’s open source.

photo provided by interviewee

the following is a summary of the conversation between 36kr and csdn chairman jiang tao:

36kr: you can first briefly introduce the community products released this time.

jiang tao: the ai ​​model community product released this time is actually a large model community that benchmarks huggingface. now the developer core needs two bases, one for the code and the second for the data and models.

in the past, huggingface was mostly used as the base for data and models. huggingface can no longer be used, so we need to construct a base equivalent to all types. its technical logic is the same.

software needs to take on a new form, and now we have seen some signs and dimensions. for example, the de-appization of mobile phones is now a transitional stage.

for example, the cross-app capabilities demonstrated by honor this time. for example, you have subscribed to many monthly subscription services, but you don’t know which ones they are, and sometimes you forget them. if you say to the honor phone, "find out all my monthly services," it will understand your words, and the mobile operating system can control your app and simulate a user to find out everything and make a list.

therefore, the app only provides capabilities and does not provide an interface.

that was on the user side, then on the development side, the tooling side, and now that's changing as well. now programmers need a tool platform to assemble modular code, because the assembly process is also a data process, which requires us to have our own base and tools, so now we have done this model and data with huawei cloud base.

we will also release an independent ide tool later. overall, we hope to do three things, which are also the core issues of open source: sustainable open source base and root ecology, trustworthy open source libraries, and intelligent software development process tools.

36kr: a year ago, domestic ai programming tools were still blank. nowadays, domestic manufacturers’ large models basically provide ai programming tool capabilities. what is the production relationship between gitcode and its ai model community and these programming tools? are they in competition with each other?

jiang tao:not counting competition. generating code for large models is an ability. it can be considered that all large models have this ability, buthow to use this capability into its processes and coding tools to make it easier to use is what we need to do now.

you have to use coding tools. all these things will end up being used by you. if you use a large model to generate code and then copy it and compile it, the efficiency is actually relatively low.

but now we have integrated it in one stop, and the code you write can be run immediately. after you finish writing the code, your program will come out immediately. then you call the gpu behind the large model, and we can also turn it into one for you according to the scheduling capabilities. keys can be called, which makes it more convenient.

36kr: how are the current batch of ai programming tools that have appeared in china different from the products on the american market?

jiang tao:after llama3 was announced, i think the essential gap between china and the united states is not too big.the key is what the gap will be like in the next generation.

now let’s talk about the logical chain, and then chatgpt o1 comes out, so o1 is a line, and the other line is not made with transformer, but with a new model, similar to li feifei’s spatial model. another thing is to prepare to add a layer on top to converge your tasks.

so now the gap between us and the other side is here. it does not lie in the announced capabilities. the gap lies in not knowing how much the other side has in reserve.at present, it seems that the follow-up gap may be larger than we imagined, because musk has reserved 100,000 chips, which shows that there must be corresponding demand.

but these large model manufacturers in china may have 20,000 to 30,000 pieces, which is about the same level.
 

36kr: the ai ​​programming tool market in the united states is relatively fragmented, with a large number of startups. how do you infer the future competition shape of this market in china?

jiang tao:the country is already relatively concentrated. the situation in china is quite special, because there is no such market in china, and most of the people who started saas in the last wave failed, so now everyone is very cautious.

now we just want to create such an environment for everyone,that is to say, we also have to build a supply chain, and what we actually build is a basic supply chain. there are still many points in the intermediate links that can be improved and improved, and everyone must be able to receive money from these points.

what we need to establish now is a commercialization path, and we must be able to earn income by making development tools.an easier path now is to start a second venture based on what you have done abroad, first make plug-ins and provide services, and then you can make some money abroad and look for the next opportunity in china.

a large number of listed software companies in china do not make money for this reason. because they have no product capabilities, only integration capabilities, they cannot make the best money for software, which is the money from copying.

36kr: how is gitcode’s commercialization path planned?

jiang tao:commercialization is actually clear now, it just needs to be done well.

first, developers are willing to pay for tools if you can help them improve efficiency.in fact, the most willing to pay is the business owner, because the boss buys developers just to buy his work efficiency. but now this consensus or common sense is not accepted by everyone, so we need to prove this.

this is the first one. we must first cut into ai programming. after cutting in, everyone found that there is actually room for improvement in every link. so, one is to find other people's open source software to assemble and use it yourself, and the other is to buy someone's ability to provide services, which is more efficient than assembling it yourself.

then we will build a centralized demand market on top of this. because we can discover some of the needs of users, we can turn this demand into a service and let developers do it. this is what we are doing next. but we have to wait until next year to see how it goes.

this requirement has been raised. if it is a requirement that can be standardized, productized, and copied, we will encourage developers to do it, and then we can become its distributor.

this is the so-called "programmer master" concept in the future.

36kr: regarding this point, you said that a large number of independent developers will be born, like up owners. this is the supply side. how do you confirm that there is corresponding large demand in the market to consume this supply?

jiang tao:of course the requirements also need to be activated. because some requirements may not be fulfilled, but most people don’t know where the requirements are because no one collects them.what we are collecting now are developers' tool requirements. in the future, there may be application requirements, which will be reflected in the application requirements (c-side requirements), but there must be a two-way and two-wheel rolling process.

we also have a sponsor called micro-saas union. isn’t it difficult to start a business now? but big companies are still optimizing their personnel. this group of people who came out think that it is better to work in the overseas market. if they find demand points overseas, if they can sell it, many independent developers can also make a lot of money. the united states itself has this software. it’s a paid market, so everyone should practice their skills there first.

in china, there is still a process to go through.this process is that the cost of software creation really needs to be reduced.our tool chain even includes foreign tool chains, and everyone is not using it so smoothly. if it can be reduced, the demand may be met.

but what are the needs now?nowadays, they are all enterprise-level requirements.the complexity and delivery standards of enterprise-level requirements are relatively high and have not yet reached the requirements of retail investors, so this is a two-way process.

this is why i said that we can have opportunities to cooperate with mobile phone manufacturers in the future. for example, there may be a lot of demand for siri.there must be a lot of people shouting about apple, but no one has refined it.

this is a point that must be discovered by programmers themselves, but it turns out that the paid channel has not been activated, so programmers may have made hundreds of thousands of applications and users will not call them.

so what we need to look at next is to find a collision point between "diversification of demand" and "low-price supply".

we recently recruited a young man to make his own ide. he was originally a core engineer of a game engine, and later started his own business. he made his own ide to help python developers create applications.

the most people learn python, but python has a huge problem. it cannot write an independent application. the process is very complicated because there is no such demand abroad. foreign professional programmers do not use python. they only need to call it when they need it. when calling some model data,

a while ago in china, artificial intelligence taught a group of people how to use python, but they couldn't make applications. (the requirements) were all scattered there. even though they knew the grammar, they couldn't make it. he saw this need and made a python ide.

then after finishing this thing, the tool was written, but no one used it, or people were not very good at using it, so he wanted to make a sample program. what kind of board program can be written to make people interested? he went to xianyu and taobao to look.

so this requires a process,

36kr: this is equivalent to the role of the gitcode ai model community, right? gather the needs of retail investors and then connect developers with the needs.

jiang tao: yes. there are no programmers on taobao xianyu, but many people send requests there. moreover, if i need to re-do the requirements that i have done, the cost will be low, because 90% of my programs can be reused.

follow for more information